Learning a variable-clustering strategy for octagon from labeled data generated by a static analysis

摘要: We present a method for automatically learning an effective strategy clustering variables the Octagon analysis from given codebase. This learned works as preprocessor of Octagon. Given program to be analyzed, is first applied and clusters in it. then run partial variant that tracks relationships among within same cluster, but not across different clusters. The notable aspect our although based on supervised learning, it does require manually-labeled data. ask human indicate which pairs codebase should tracked. Instead uses impact pre-analysis previous work labels variable positive or negative. implemented top static buffer-overflow detector C programs tested against open source benchmarks. Our experiments show with scales up 100KLOC 33x faster than one (which itself significantly original analysis), while increasing false alarms by only 2 %.